Compressing Multidimensional Weather and Climate Data Into Neural Networks

Learn about a novel method for compressing multidimensional weather and climate data using neural networks in this 16-minute conference talk. Explore how coordinate-based neural networks can be trained to overfit data, resulting in a compact representation of grid-based information. Discover the impressive compression ratios achieved, ranging from 300x to over 3,000x, and how this approach outperforms state-of-the-art compressors in terms of weighted RMSE and MAE. Understand the method's ability to preserve important large-scale atmospheric structures without introducing artifacts. Examine the practical applications of this compression technique, including its use as a 790x compressed dataloader for training weather forecasting models with minimal impact on accuracy. Gain insights into how this groundbreaking approach can democratize access to high-resolution climate data and open up new research possibilities in the field of weather and climate science.

Background
Introduction
Results
Applications
Conclusion